The market doesn't always reward you for seeing the future early. It rewards you when everyone else finally sees what was already happening and you held on to your marshmellow. But holding onto one marshmellow to get the second one a year or two later isn't always easy. In 2023–24, my $GOOGL thesis was vertical integration. Google had spent ~10 years developing TPUs, while also owning the AI model, cloud infrastructure, Search, YouTube, Maps, Gmail and distribution to billions of users. The pieces were already there. The market just didn’t fully appreciate them yet. I had $100K in $GOOGL LEAPS every year from 2023 to 2025. I cashed out when $GOOGL was at $207 a share. I wanted back in but my attempts to buy $100K on multiple days didn't fully execute. Instead I watched $GOOGL rise to $300+ and missed out on $700K in gains. I ate one marshmellow and watched others eat the second one. In 2024, I saw the same kind of asymmetry in $PLTR: best-in-class pattern recognition applied to defense and commercial datasets to increase revenue, improve efficiency and reduce costs. A perfect balance sheet: no debt, $5B in cash, no dividend payout. I turned roughly $20K in LEAPS into $500K, then converted part of that gain into real estate. I bought a house for cash in 2025 at what I believed was a >$100K margin of safety, refinanced in 2026, and bought a $600K duplex. Today, roughly $2K/month goes toward principal paydown while the duplex produces positive cash flow. The real estate also gives me a hedge on inflation (and my LEAPS investments) as well as ~$30K/year of depreciation that offsets rental income elsewhere to decrease my tax burden. $PLTR wasn’t just a stock win. It became productive assets and long-term tax savings. In 2024, I put $40K into $AMD LEAPS. Simple thesis: The GPU pie was growing, and $AMD was going to grow with it as its product cycle improved, allowing its margins to improve. I had big $PLTR capital gains at the end of 2024 and sold $AMD at a loss to reduce my tax burden. If I had held on the marshmellow, I would have eaten a second one in 2026. Stock was better than LEAPS for $AMD. Last year I saw something similar with $META. Zuckerberg wasn’t spending tens of billions on AI infrastructure CapEx just to recreate the Metaverse experience. His jiujitsu training made him more focused. $META doesn’t want to permanently pay $NVDA margins for chips or $AMZN/$MSFT/$GOOGL margins for cloud. I had $50K to $100K in $META in 2025-2026. I sold $META in June 2026 and bought $TSM and $AMD. I gained about $8K this week on the ~$45K rotation. I wanted back in on $META. No money, no buy (the original version of "No woman, no cry). In 2025, $MRVL was a similar thesis as $AMD in 2024: The #2 player in one field, #1 player in another priced well below its value. I bought $31K in LEAPS which are now worth $700K+ in unrealized gains. My plan is to exercise most of the position, defer the taxable event where applicable, use the shares as collateral for portfolio-covered puts, and recycle that premium into LEAPS. $PLTR made my real estate business more profitable. $MRVL will help me create a new options "business" to generate about $30K+ in gains per year selling puts while I hold $1M in $MRVL stock waiting for it to 4x per Jensen Huang's $1T thesis. And this is why I keep coming back to the marshmallow test. Most investors don’t want to buy while a company is still investing and the outcome is uncertain. They want certainty. But certainty is expensive. They buy $PLTR at $100+ instead of $7-$25. They buy $AMD at $300-$650 instead of $75-$200. They buy $MRVL at $200+ instead of $65-$100. They buy $GOOGL at $250–$400 instead of $150–$200. They buy $META at $650–$750 instead of $520–$600. By the time everyone can clearly see the future, the market has usually started charging admission. Everyone wants the marshmallow. Few know how to get two.

$MRVL’s Oct. 6 Investor Day could be a major rerating event. The key is whether Marvell can convince investors that AI growth extends well beyond FY28. The market already has roughly $12B FY27 and $18B FY28 revenue expectations. What it doesn’t have is a clear FY29–FY30 bridge. If Marvell lays out a credible path toward $40B+ FY30 revenue, the valuation framework changes. Then there’s 2nm. Marvell already has a 2nm platform for next-generation custom XPUs and networking, 2nm custom SRAM, 2nm chiplet interconnects, 2nm coherent DSPs and 400G/lane optical technology. The big question is how many hyperscaler programs are moving onto that platform and when they enter production. Add the $GOOGL custom silicon relationship, custom HBM, scale-up switching, CPO, and Celestial AI, and Marvell increasingly looks less like a collection of semiconductor products and more like an end-to-end AI infrastructure platform. That matters because a rerating can come from two directions at once: - Higher long-term EPS estimates - Higher valuation multiple What I’ll be watching Oct. 6 are FY30 revenue, FY30 margins/EPS, 2nm design wins, custom silicon growth, scale-up TAM, and the Google ramp. If those surprise to the upside, $MRVL may not just get higher estimates. It could get a different multiple.

Jensen Huang and $NVDA making the right moves.

Let's revisit that interesting close last Friday (9/18) to see how stocks closed this Friday (9/25)... Stocks that closed above resistence on 9/18: $NBIS > $220 on 9/18 vs $237 on 9/25 $GLW > $150 vs $157 $MRVL > $240 vs $262 $ALAB > $300 vs $365 $NVDA > $220 vs $225 $TSM > $430 vs $450 $MU > $1000 vs $1082 $SPY > $760 vs $771 $QQQ > $720 vs $745 Result: All higher. Stocks that closed below resistence on 9/18: $ANET < $200 on 9/18 vs $207 on 9/25 $HOOD < $120 vs $119 $VRT < $250 vs $253 $META < $670 vs $752 $AMD < $560 vs $631 $SKHY < $190 vs $192 $BA < $200 vs $198 $V < $370 vs $367 $CSCO < $110 vs $107 $ASML < $1680 vs $1744 $AVGO < $359 vs $353 $ORCL < $149 vs $137 Result: 6 down. 3 up a tad. 3 up significantly (Meta trade). Take away the Meta trade and these stayed near or below resistence from 9/18. On 9/18, some stocks rose and closed just above resistance but fell below in after-hours, showing potential weakness: $LITE < $930 on 9/18 vs $942 on 9/25 $GEV < $940 vs $958 $AAOI < $105 vs $101 $SOFI < $17.0 vs $16.6 Result: Half up, half down. On 9/18, other stocks rose and were below resistance at closing but worked their way above resistance in after-hours, showing potential strength for next week: $GOOGL > $350 on 9/18 vs $344 on 9/25 $INTC > $109 vs $123 $VO > $80 vs $80 Result: Mixed.

Growth everywhere. $TSM, $MU, $SKHY, Samsung and other $DRAM companies expanding production. $MAGS CapEx expanding. EPS of $SPY companies growing. New apps and startups beginning to make use of AI. Muse, OpenAI, and Anthropic more useful to small businesses and sole proprietors. New waves of AI adoption in other countries and they invest in AI. Yet multiples of many companies remain low because analysts lack historical perspective and visibility? $NVDA, $MU, $SKHY and others will be rerated within two years. Stock prices will keep going up. The S&P 500 will hit 8000 in 2026 and probably 10,000 in 2027 and 12,000 in 2028. Those who get it say we’re in the early stages of a transformational period of growth. Granted, an innovation could make everything too cheap or debt problems or war could derail everything for a moment. But the train would simply be put back on the track with a minor delay. This change is inevitable by now, whether we like it or not. The elite and economically powerful have decided for us. Old, tech-stupid politicians aren’t stopping it — especially when it’s framed as “us vs them” or “China is winning” fear-mongering. Consumers will buy the new gadgets they always buy and AI will be included. Adoption will happen as an afterthought — to hell with negative effects on human behavior, the environment, jobs, or other aspects of life. Those who invest can soak up profits, pay your govt capital gains taxes, and hope the money is spent on something better than bombs or debt interest payments. It’s the invisible hand working to block our eyes from seeing everything clearly. But at least we can become millionaires and create generational wealth for our kids who will live in a strange new world of robots, flying cars, food and package deliveries by air, and other mostly unnecessary innovations.

$META unveiled a new product. Cool. I’m bullish on $META, and have been since last year when the CapEx was clearly (to me) showing how $META was vertically integrating (chips/cloud telegraphed in 2025, AI model announced in early 2026) and financials showed how using AI translated into revenue growth. But exaggerated excitement about VR glasses? Come on! 1. Isn’t everyone critical of the Metaverse? What is VR if not a major aspect of that? Zuck is repackaging the Metaverse failure. 2. $META’s existing glasses are cool. At some point I will buy them to visit properties and record what I see. This will add value to my real estate investing. But they aren’t a big hit. Think about it. For millions of years, humans have an advantage over other animals in part due to the dexterity of our hands. We want to hold and manipulate objects with our hands. It’s inherent to our being, a truly primitive urge. Similarly, for millions of years we prefer unobstructed vision. We dislike things on our faces or covering our eyes. If you dislike the evolutionary argument for why $GOOGL Glass failed and why $AAPL and $META AI glasses don’t catch on, how about a simpler idea: geeks wear glasses. Isn’t that what we’ve been taught by American culture? How many people with exaggerated excitement about Meta’s VR glasses wear contact lenses? Why don’t you wear normal glasses? They’re more convenient than sticking something onto your eyeball. It’s cool to manipulate objects with your hands (think smokers trying to act cool or people taking out their phones in front of others to act cool). It’s not cool to wear glasses. If you disagree, ditch your contacts and buy Meta VR glasses in the same week. Lastly, 90% of people expressing exaggerated excitement about VR glasses will buy 3-10 Apple products (phones or wearables) before they buy a pair of VR glasses. It’s simply against our nature and not cool. Refocus your attention on what will move $META EPS: Marketplace getting its own app and bringing in money; Meta’s in-app dating service making more than dating apps; Meta Cloud being bigger than $CRWV, $NBIS and $IREN; Meta chips and cloud and Muse reducing long-term OpEx and increasing revenues because they will be designed specifically for $META products and services. $META doesn’t create slick, cool videos about what it is telegraphing. You need human intelligence to put the pieces together.

The WSJ also uses construction, financing/risk, jobs, the wealth effect, and inflation to argue that “The AI Build-Out Is Becoming the Biggest Economic Bet in U.S. History.” Unfortunately, there is no historical comparison. The title of the article (above) states a historical argument — it’s “becoming the biggest economic bet in US history.” Well, a single chart lacking context in the beginning of the article doesn’t suffice if the 5 main categories (listed above) supporting the argument make no historical comparisons. Construction of railroads took at least 100 years, from the 1820s to the 1920s and beyond. That’s not a “bigger bet”? Financing and risk connected to the railway buildout caused the panics of 1857, 1873, and 1893. In the 1800s, the financial system (banking, stock market) was unregulated and messy, while rampant political corruption greased the wheels of business. Will the AI Boom give way to an AI Bust? Currently, it seems the AI Boom avoided a recession in ~2022-2024. The ChatGPT moment of late 2022 coincides with the incessant use of “K-shaped economy”. Jobs associated with electrification aren’t potentially more significant? If it’s dark, nobody works. Now, AI will add some jobs, replace possibly more jobs, and allow many (unmotivated) employees to use AI to do their work even though they could have done it themselves. I’d argue that electrification led to more job growth than AI will. Railroads transformed transportation of agricultural, manufacturing, and consumer goods. The mighty American export machine that led to American imperialism in search of new materials and markets would have never happened without railroads. The Second Industrial Revolution stalls without railroads. The wealth effect and inflation are nothing new. In transformational times, wealth flows to the owners of the new technologies. The Gilded Age was also marked by great inequality. Yet industrialization led to higher literacy rates, more economic power for women, more social mobility, and a higher overall standard of living for everyone — even if the rich got substantially richer. New technologies and infrastructure investments also led to worse inflation in previous generations. Why? First, inflation wasn’t properly measured or really understood. You can’t solve a problem whose cause-effect relationships aren’t part of public policy discourse. Second, the financial system was less efficient, governments were more laissez-faire, the income tax didn’t exist and (contrary to Trump’s tariff talk) tariffs didn’t bring in enough money to support public policies to help people in need. In previous periods of panics, busts, or recessions, a market crash was utterly devastating for lack of a safety net and effective govt. If the Covid recession never happened and the Fed’s financial engineering never happened, would govt debt across the world be so high? Europe doesn’t import $ASML lithography machines at all — meaning they don’t create semiconductor chips — and EU countries lag substantially in terms of data-center buildout. Yet their bond prices are also reaching highs and the European Central Bank increased rates. Their inflation situation has arguably been worse since the AI buildout began, though that is due to greater economic interdependence with Russia and Iran, which have been disrupted by war. The inflation argument in the WSJ article is questionable. In the 2010s, the Fed preferred more inflation. Remove Covid and war from the equation and maybe the AI buildout would have been completely manageable, unlike previous eras marked by boom-bust cycles.

The WSJ uses % of GDP as the core indicator to argue that “The AI Build-Out Is Becoming the Biggest Economic Bet in U.S. History.” Let’s add some contextualization. 1. The data shown isn’t the full story. The AI buildout started earlier. The telecom and fiber buildout is part of the AI buildout. The railroad buildout started in the 1820s or earlier. A major aspect of New Deal policy was Hoover Dam and Tennessee Valley Authority in the 1930s. The electrification and RR buildouts cover 20 years each. Will historians in 100-200 years simply conflate the telecom, fiber, cloud, and AI buildouts as one period, from c. 1996 to 2032? Or will they categorize the buildouts into two groups — Internet/telecom (c. 1990-2005) and smartphones/cloud/AI (c. 2005-2035)? 2. The railroad buildout included federal and state land grants from 1850 to 1871 of about 180 million acres, or about 10% of the continental United States. These grants were used to incentivize the RR companies to build. The WSJ conveniently doesn’t include 1850-1870 in its data. When govt land was granted to RR companies, how much was the land worth? Not so much. Once the railroads were built, RR companies sold some land to speculators and kept other lands for their stations and to sell at a later time at a higher price. Combine that with the concomitant settling of the West from the time gold was discovered in California immediately following the Mexican-American War (1846-1848) and the transfer of ownership of another 275 million acres (~14% of the total land of the continental US) following the Homestead Act of 1863, and the historical significance of the railroads becomes far more interesting. Millions of people aren’t moving across the country to work near a data center today. 3. Different times, different labor costs. The data-center buildout involves specialist labor — tradesmen, tech companies. Of course that is more expensive than the cost of abundant cheap immigrant and slave labor. Immigrants from Latin America today (the modern-day equivalent of exploited labor) are not driving the AI buildout. Let’s compare differing versions of “Chinese labor” in the 1800s and now: during the railroad buildout the Chinese (residents, eventual citizens, and citizens) were the cheapest non-slave labor; now, engineers of Chinese origin (residents and citizens) are some of the most expensive workers, especially if they work for $NVDA, $AMD, $AVGO, $MRVL, $MU and other semiconductor companies.

$AMD has a dilution issue worth watching. OpenAI can buy up to 160M AMD shares for $0.01 each. $META can buy another 160M. Total: 320M new shares. AMD currently has ~1.63B shares outstanding, so full exercise would increase the share count by roughly 20%, diluting existing shareholders by about 16%. But this won’t happen overnight. Both deals require AMD GPU purchases to scale from 1 GW → 6 GW, plus stock-price and performance milestones. My rough timeline: Late 2026–2027: first warrant tranches could become exercisable as the first 1 GW deployments arrive. 2027–2030: additional tranches vest as Meta/OpenAI scale purchases. By 2030–31: potentially the full 320M shares. Important: this does NOT mean AMD suddenly drops 16% when the warrants are exercised. The market should gradually price in the dilution as the warrants vest. But mathematically, if AMD’s market cap stayed unchanged, full dilution would make each share worth about 16% less than it otherwise would have been. The bet AMD is making is simple: Give Meta + OpenAI ~$200B of potential equity at today’s stock price… …in exchange for turning them into massive, multi-generation AMD GPU customers. If the resulting earnings grow faster than the share count, the deals can still be very bullish for $AMD.

2/ I don’t think advertising inside Pinterest is the endpoint of its revenue-generation strategy. ethereum:0x2e44f3f609ff5aa4819b323fd74690f07c3607c4 has already announced an MCP server designed to bring Pinterest intelligence directly into AI assistants, copilots, and agentic workflows. That may be the clue to the much larger opportunity. My prediction is that $PINS eventually monetizes the Taste Graph in three layers: 1. Closed-loop commerce (in $PINS's website and app) Merchants upload catalogs. Pinterest AI predicts what users want. Performance+ and Visual Search Ads monetize the recommendation. Conversion data comes back and improves the model. This makes Pinterest increasingly resemble a visual Google Search + performance advertising engine. 2. Commerce Intelligence $WMT, $TGT, brands, and retailers could eventually pay Pinterest not necessarily for the raw Taste Graph, but for intelligence produced by it: What styles are emerging? What products belong together? What does this customer segment want next? What visual attributes are driving purchases? Pinterest keeps the proprietary ontology while selling increasingly valuable outputs from it. 3. AI-agent distribution This could be the biggest opportunity. Instead of requiring someone to open Pinterest, imagine: ChatGPT, $GOOGL Gemini, $MSFT Bing, $AMZN shopping agents or other AI systems ask Pinterest: “Find furniture matching this room.” “What products fit this person’s style?” “What is trending among women planning weddings?” Pinterest’s MCP/API infrastructure could allow external AI agents to query its Taste Graph and visual-search intelligence. That would remove one of Pinterest’s biggest limitations today: Pinterest currently monetizes primarily when its own users interact with Pinterest. An AI-agent model could allow Pinterest to monetize its intelligence when billions of people search and shop outside Pinterest too. And importantly, Pinterest wouldn’t need to sell its underlying data. It could sell the differentiated answer. That fits Linares’ framework almost perfectly. Generic AI models become cheaper. But the token: “Here are five couches” is not worth the same as: “Based on years of your visual preferences, the products people with similar taste actually purchased, current merchant inventory and emerging style trends, these are the five couches you are most likely to buy.” The model may be commodity. The Taste Ontology is not. Pinterest’s missing piece is still the outcome side of the loop. From Linaers' framework, $SPOT knows whether you played or skipped the song. $DUOL knows whether your answer was correct. Pinterest knows enormous amounts about intent, but it doesn’t always know whether the user completed the purchase. Closing the loop is essential. Visual Search Ads, merchant catalogs, Performance+, conversion APIs and deeper commerce integrations appear to be moving directly toward fixing that. That is why I think 2027 could be the important year for $PINS. 2026 is the infrastructure/product-build year. 2027 could be when we start seeing whether higher Ontology Velocity translates into higher advertiser ROAS, ARPU, ad pricing, margins and EPS as ethereum:0x2e44f3f609ff5aa4819b323fd74690f07c3607c4 closes the loop. If that happens, the market may eventually stop viewing Pinterest as a social-media advertising company. It may start viewing it as something much more interesting: the commercial ontology of human taste. I predict that the majority of $PINS's future revenues by 2030 will be related to its AI strategy today.

1/ I think the market may be misunderstanding what $PINS is building with AI. Antonio Linares’ “Ontology Velocity” framework argues that winners in AI won’t necessarily be the companies with the biggest models. Models are becoming cheaper and increasingly commoditized. The moat is the proprietary closed loop: behavior → intervention → outcome → feedback → better next intervention. That produces what Linares calls non-replicable tokens: AI outputs that are more valuable because they are informed by proprietary data competitors cannot reproduce. $PINS fits this framework surprisingly well. It has 640M users telling it what they like through searches, saves, boards, clicks, visual searches, and product interactions. A user doesn’t just look at a chair. They save the chair to “Living Room Ideas.” They save a walnut table next to it. They search for warm lighting. They ignore one recommendation and click another. Hundreds of millions of people have been organizing the visual internet this way for years. That is the Taste Graph. And Pinterest is now spending aggressively to turn that ontology into an AI system. $PINS committed $4B to $AMZN AWS through 2031 to expand AI compute. It built a new multimodal AI infrastructure layer with $NVDA. It launched Pinterest Assistant and Ask Pinterest. It launched AI-powered Performance+ advertising. And this month it launched Visual Search Ads, allowing merchants to bid for placement precisely when users are visually searching, comparing products and moving toward a purchase. Pinterest says users conduct 80B+ searches per month, most visual, with more than half carrying commercial intent. This is where the strategy starts becoming clear. Pinterest already knows: What do you like? The next question is: What will you buy? Closing that loop would dramatically increase Pinterest’s Ontology Velocity. Taste → recommendation → merchant ad → click → purchase → conversion data → better prediction. Every successful transaction then becomes another labeled training event. $PINS then has differentiated, non-replicable tokens that not only generate income within the app and website of Pinterest. Others could also benefit from the highly differentiated visual data of Pinterest. Think $AMZN, OpenAI, Anthropic, $TGT, $WMT.

Sold $TSM to buy $GOOG. I always buy $GOOGL but I noticed there is a $5 discrepancy between the two. Usually they are closer than $5. I think Alphabet has more upside than TSMC. The $5 difference between $GOOGL and $GOOG is an added bonus.

$VST is starting to resemble its previous bottoming setup. Last time, the stock formed a higher low, reclaimed its moving averages, then broke sharply higher. Now VST is again trying to build a higher low around $132–140. Key confirmation zone: $146–156 — reclaiming the 50/100/200-day MAs would make the comparison much stronger.

$230 has been the Iron Curtain for $NVDA. If NVIDIA can break through and close above $230 for multiple days, $230 might serve as a formidable level of support in the future. That implies that $NVDA would finally be off to the races to $240, $250, and above .

Is a deal with Iran near? Trump met with the Saudi crown prince at the White House last week. Today or tomorrow Trump might meet Iran’s president at the UN. Then on Thursday Trump is expected to host Xi Jinping at the White House (assuming Xi is well). China could use its influence to help get Iran’s extremists to the finish line of negotiations.